Cuttlefish: A Scaling-Aware Adapter for Structured Reasoning

The ICML 2026 accepted paper Cuttlefish introduces a novel scaling-aware adapter aimed at significantly enhancing the structured reasoning capabilities of large language models (LLMs) without modifying the base model parameters. This adapter features dynamic adjustment of reasoning depth and width, can be transferred across models of different scales, and achieves a balance between computational efficiency and reasoning quality.

Background: Structural Challenges in Large Model Reasoning

In recent years, large language models (LLMs) have demonstrated remarkable capabilities across various tasks, but when faced with complex problems requiring multi-step logical reasoning, models often suffer from "hallucinations" or broken reasoning chains. While traditional fine-tuning methods can improve performance on specific tasks, they struggle to fundamentally enhance the structured reasoning capabilities of models. Structured reasoning requires models to organize their thinking process in a logical sequence, with each step built upon the previous one. This ability is crucial for tasks such as mathematical proof, code generation, and scientific reasoning. However, many existing methods either have excessively high computational costs or are difficult to transfer across models of different scales.

Core Innovations of Cuttlefish

Cuttlefish is a novel scaling-aware adapter, whose name is inspired by the complex nervous system and environmental adaptability of cuttlefish.

Scaling-Aware Mechanism

The core innovation of Cuttlefish lies in its "scaling-aware" feature: it dynamically adjusts the depth and width of reasoning based on the complexity of the input problem—using shallower paths for simple problems and activating deeper modules for complex tasks. This not only maintains quality but also reduces average computational costs, and is compatible with base models of different scales (from 7B to 70B+).

Structure-Grounded Reasoning Framework

A structure-grounded reasoning framework is introduced, which requires building a complete reasoning graph before generating answers. It consists of three components:

1. Reasoning Node Recognizer: Automatically identifies key reasoning nodes
2. Edge Relationship Modeler: Establishes logical dependencies between nodes
3. Path Validator: Verifies the integrity of the reasoning chain

Technical Implementation Details

The technical architecture adopts a lightweight parallel adapter design, inserting a learnable structure adaptation module between the attention and feed-forward networks of the Transformer layer, which interacts with the original model through a gating mechanism. Training uses a multi-task learning strategy to optimize reasoning accuracy, structural integrity, and computational efficiency. Training data is constructed via an automated reasoning path annotation process, extracting high-quality structured samples from mathematics, logic, and code datasets.

Experimental Results and Performance

In the ICML 2026 review process, Cuttlefish outperformed others on multiple structured reasoning benchmark tests. Especially in mathematical problem-solving tasks requiring long-chain reasoning, it achieved an accuracy improvement of 15-25 percentage points compared to traditional prompt engineering methods.

Its cross-scale generalization ability is outstanding: the adapter trained on a 7B model can be applied to 13B and 70B models with minimal adjustments, with a performance loss of no more than 5%, which is of great significance for practical deployment.

Practical Application Value

Cuttlefish provides a new tool for LLM application developers, suitable for scenarios requiring reliable reasoning such as educational tutoring, scientific research assistance, and code review. It can improve output quality without replacing existing models.

Its lightweight design is suitable for resource-constrained environments, with an increase in reasoning overhead usually not exceeding 10-15%, allowing it to run smoothly on consumer-grade hardware.

Summary and Outlook

Cuttlefish represents an important advancement in the field of LLM reasoning enhancement. Through its scaling-aware and structure-grounded design, it balances computational efficiency, reasoning quality, and model generality.

As LLMs are increasingly applied across various industries, the demand for reliable reasoning continues to grow. Technologies like Cuttlefish are expected to become standard components of next-generation AI systems. The research team has open-sourced the code and pre-trained weights, and looks forward to the community further exploring the boundaries of reasoning enhancement.